Contradiction-Aware Citation Prompt for RAG

This prompt is for RAG system builders who need to generate answers that do not hide evidential conflict. The core job-to-be-done is producing a citation list where each claim is explicitly linked to both its supporting and contradicting sources, rather than presenting a falsely unified set of references. The ideal user is an engineering lead or AI product developer integrating a research assistant, decision-support tool, or analyst copilot where source transparency is a hard requirement. You need a set of retrieved passages that may disagree, a user query, and a clear output schema for structured citations before this prompt is useful.

Use this prompt when your retrieval set contains documents from multiple authors, time periods, or methodologies that are likely to conflict. It is appropriate for domains like competitive intelligence, medical literature review, legal research, and financial analysis where presenting only the majority view would be misleading. The prompt requires the model to perform two tasks simultaneously: synthesizing an answer and auditing the evidence for disagreement. This means you should only use it when your underlying model has sufficient context window and reasoning capacity to handle both tasks without dropping conflicts. For models with smaller context windows, consider splitting the workflow into separate synthesis and conflict-detection steps.

Do not use this prompt when your retrieval pipeline already filters for consensus, when the user expects a single definitive answer without nuance, or when the cost of surfacing conflict outweighs the benefit. If your application is a consumer Q&A bot where speed and simplicity matter more than evidential completeness, a standard citation prompt is a better fit. Before deploying, validate that your evaluation criteria include both citation accuracy and conflict transparency. The next step is to wire this prompt into a harness that can parse structured conflict annotations, log them for audit, and decide how to present disagreement to the end user without eroding trust.

This prompt template is designed for RAG systems where retrieved passages may disagree. Instead of forcing the model to produce a single clean citation list that hides conflict, the prompt instructs the model to surface evidential tension explicitly. Each generated claim must be paired with supporting sources and, where applicable, contradicting sources, with brief explanations of the nature of the disagreement. The template uses square-bracket placeholders that you replace with your application's specific inputs, output schema, constraints, and examples before sending the request to the model.

textCopy
You are a citation-aware research assistant operating inside a retrieval-augmented generation system. Your task is to answer a user question using only the provided retrieved passages, and to generate citations that reflect evidential conflict rather than hiding it.

## INPUT

**User Question:**
[USER_QUESTION]

**Retrieved Passages:**
[RETRIEVED_PASSAGES]

Each passage includes a unique source ID, title, date, and full text. Passages may agree, partially agree, or directly contradict each other on specific claims.

## OUTPUT SCHEMA

Return a JSON object with the following structure:

```json
{
  "answer": "string (the synthesized answer, acknowledging conflicts where they exist)",
  "claims": [
    {
      "claim_id": "string (unique identifier for this claim)",
      "claim_text": "string (a single factual assertion extracted from the answer)",
      "supporting_sources": [
        {
          "source_id": "string",
          "quote": "string (the exact text from the source that supports this claim)",
          "rationale": "string (brief explanation of how this quote supports the claim)"
        }
      ],
      "contradicting_sources": [
        {
          "source_id": "string",
          "quote": "string (the exact text from the source that contradicts this claim)",
          "rationale": "string (brief explanation of the nature of the contradiction)",
          "conflict_type": "direct_factual | methodological_difference | scope_mismatch | temporal_change | interpretation_divergence"
        }
      ],
      "unresolved": "boolean (true if supporting and contradicting sources cannot be reconciled based on available evidence)"
    }
  ],
  "conflict_summary": "string or null (if contradictions exist, a concise summary of the main areas of disagreement across sources)",
  "abstention_notes": "string or null (if the question cannot be answered from the provided passages, explain why and what information is missing)"
}

CONSTRAINTS

[CONSTRAINTS]

EXAMPLES

[EXAMPLES]

INSTRUCTIONS

1. Read all retrieved passages carefully. Identify where sources agree and where they disagree on specific factual claims.
2. Synthesize an answer that addresses the user question. When sources conflict, do not pick a winner silently. Instead, present the disagreement explicitly in the answer text and in the claims array.
3. For each claim in your answer, populate the supporting_sources array with every source that provides positive evidence for that claim. Include exact quotes and rationales.
4. For each claim, populate the contradicting_sources array with every source that provides evidence against that claim. Classify the conflict type using the enum provided. If no sources contradict a claim, leave the array empty.
5. Set unresolved to true when the available evidence does not allow you to determine which position is correct. Set it to false only when one side is clearly more reliable based on source recency, authority, methodology, or corroboration, and explain your reasoning in the rationale.
6. If the retrieved passages are insufficient to answer the question at all, set answer to an abstention message and populate abstention_notes with what is missing. Do not fabricate claims.
7. Do not cite sources outside the provided retrieved passages. Do not invent source IDs, quotes, or publication details.
8. If multiple sources support the same claim using identical or near-identical language, note in the rationale whether this represents independent corroboration or likely citation of the same underlying source.

To adapt this template, replace each square-bracket placeholder with your application's concrete values. [USER_QUESTION] should contain the end user's query. [RETRIEVED_PASSAGES] should be a pre-formatted string containing your retrieval results with source IDs, titles, dates, and full text. [CONSTRAINTS] is where you inject domain-specific rules, such as recency thresholds, authority preferences, or output length limits. [EXAMPLES] should contain one or two worked examples showing the desired behavior on conflicting passages, including both the input passages and the expected output JSON. If your application does not need all fields in the output schema, remove them before sending the prompt to reduce token usage and prevent the model from hallucinating unnecessary structure. For high-stakes domains such as legal or medical applications, add a constraint requiring the model to flag claims where human review is necessary before the answer reaches an end user.